Codemcp

What ezyang/codemcp MCP server does

The ezyang/codemcp MCP server connects Claude to a Git repository on the local machine. It supports the basic operations needed for coding tasks: inspecting project files, writing changes, and invoking commands that the project owner has explicitly configured. This allows a developer to ask Claude to implement a feature, repair a defect, refactor code, or work through tests while keeping the repository available in the developer's normal editor.

The design emphasizes reviewable changes. Agent edits are recorded in Git, with a commit created for the work and amended as the chat continues. That gives the developer a fine-grained history for reviewing or reverting changes instead of relying on an uncommitted working tree.

How it works

The server runs locally and exposes an SSE endpoint. Claude Desktop can connect to that endpoint, and the README also describes using a browser extension that connects Claude's web interface to SSE MCP servers. A project is selected by initializing codemcp with its filesystem path.

Shell access is constrained by configuration rather than being unrestricted. The agent can only invoke commands declared in the repository's codemcp.toml file. Command entries can be simple argument arrays or can include documentation describing how a command accepts arguments. A format command has special behavior: it runs after each file edit. Other named commands, such as tests, are available for the model to choose when appropriate.

The README describes the current master branch as moving back toward multiple MCP tools, which means tool-use approval may be required. It also notes that the project is not intended to run unattended indefinitely; developers should review the result of each chat and continue with another request when needed.

Setup and configuration

Install uv, then start the server with:

uvx --from git+https://github.com/ezyang/codemcp@prod codemcp serve

The default SSE address is http://127.0.0.1:8000/sse. A different port can be supplied with --port. The server can also bind to a specified host for remote access, but exposing it broadly is unsafe because an MCP user can execute code on the computer.

Create codemcp.toml in the target Git checkout before using it. A minimal configuration might define formatter and test commands under [commands]. You can also set project_prompt, which is loaded when the project is initialized. In Claude Desktop, the README recommends a project instruction that initializes codemcp with the repository path.

Tools and capabilities

• Read project files for coding tasks.
• Write and modify files in the selected checkout.
• Run commands declared in codemcp.toml.
• Run the configured formatter after edits.
• Create and amend Git commits for agent changes.
• Write logs to ~/.codemcp/codemcp.log.

Limitations and notes

The ezyang/codemcp MCP server requires a Git checkout and project configuration. It does not provide unrestricted shell execution; required formatters, test runners, and similar commands must be declared in advance. Remote binding should be limited to a trusted network, since access can enable arbitrary code execution on the host.

The README says codemcp is considered obsolete because Claude Code is available through Anthropic subscriptions. It remains open source and may still suit users who prefer its IDE-independent workflow, Git-based edit history, or auto-accept-oriented design. Claude Pro or Claude Max is discussed as the intended subscription context, rather than a server API key requirement.